acceptodds
Withdrawn from ICLR 2027

FedCARR: Contribution-Aware Retention and Rollback Optimization for Federated Unlearning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated unlearning (FU) aims to remove the influence of clients requesting data deletion from a collaboratively trained model, but existing approaches primarily focus on the operational execution of unlearning while overlooking the economic decisions surrounding batched deletion requests. In practical federated learning systems, a server may have a limited compensation budget and must jointly determine which clients to temporarily retain and which historical model state to use for subsequent unlearning. We propose FedCARR, a two-stage mechanism for contribution-aware batched client-level FU. In the first stage, the server assigns contribution-weighted compensation caps and determines individually rational retention decisions under a hard budget constraint. These decisions and associated payments are then fixed as binding commitments. In the second stage, given the resulting set of clients to be forgotten, the server selects an executable rollback state from loadable or reconstructible historical checkpoints by jointly considering retained model utility, compensation expenditure, and delayed-deletion privacy costs. We theoretically establish individual rationality, retention-decision stability, the existence and optimality of feasible rollback decisions, and the existence of an optimal solution to the formulated problem. Extensive experiments on CIFAR-10, CIFAR-100, and Fashion-MNIST evaluate the tradeoffs among model utility, compensation cost, deletion delay, and rollback requirements, demonstrating the effectiveness of the proposed mechanism for budget-constrained FU.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Unlearning, Incentive Mechanisms, Contribution-Aware Retention, Rollback Optimization, Privacy Compensation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_n7b3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper looks at federated unlearning when several clients ask to have their data removed at once. Instead of removing all of them, the server may pay some requesters to let their data stay, and must then decide where to roll back the model for the rest. The proposed method, FedCARR, works in two steps. First, it decides which requesters to keep, based on how much each contributed and how much it would cost to compensate them. Second, it picks a single historical checkpoint from which to unlearn the remaining clients, balancing model quality against the compensation owed for delayed deletion.

Critical strengths

The problem of asking which requests to accommodate, and how that choice couples with the unlearning state, is reasonable and practical.

Critical weaknesses
  1. In the abstract and introduction, the authors mentioned the limited compensation budget and argue that retention and rollback must be decided jointly. However, this motivation is not discussed or explained carefully in the proposed method.
  2. The notations are inconsistent. For example, τi\tau_i means both the client type and round index.
  3. The experimental evaluations are not comprehensive. The experiments should be conducted on more challenging dataset. More baselines should be conducted to further prove the method's novelty. More in-depth analyses and ablation tests should be conducted to prove the paper's claims.
  4. Despite the good motivation of the rollback state, I still feel that the existence of historical states for the rollback cost a significant memory. This should be discussed very carefully of the practicality of the proposed setting.
  5. The computation and memory efficiency of FedCARR is missing. This is critical because this is one of the main motivation of the paper.
  6. The theoretical analysis is limited. Two results are incorrect as stated: the feasible rollback set can be empty, which the paper itself anticipates in Algorithm 2, and in that case problem (P1) has no solution, so Lemmas 3.2 and 3.3 need an additional assumption. The remaining results follow almost directly from the definitions and offer little insight. More importantly, the analysis omits the properties that matter for this kind of mechanism: whether clients have an incentive to report truthfully, how close the two-stage solution comes to the optimum of (P1), and whether the unlearning step actually removes the forgotten clients’ influence.
Questions for authors
  1. What is limited budget for compensation?
  2. The idea of rollback seems uncommon as current works unlearn the global model directly. The authors should discuss about the literature review more comprehensively and explain carefully.
  3. What is the meaning of Section 2.2 Unlearning Request when unlearning is mentioned in Section 2.1?
Reviewer_rD3rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies retention compensation and rollback selection for batched client-level federated unlearning and proposes a two-stage mechanism, FedCARR. Stage 1 determines the retained set and fixes payments based on clients’ historical contributions, privacy costs, and individual compensation caps. Stage 2 selects a common executable historical state for the leaving clients, balancing model value against delay compensation. The authors analyze individual rationality, leaving-decision stability, and solution existence, and evaluate compensation, retention rates, rollback selection, and unlearning performance on three image datasets.

Critical strengths
  1. The paper addresses a meaningful economic decision in federated unlearning: negotiating the retention of some requesters through compensation while coordinating the remaining deletion requests. Connecting client preferences with model maintenance provides a useful additional perspective on federated unlearning.

  2. The two-stage procedure is intuitive and straightforward to implement, fixing retention commitments before selecting a common rollback point. Experiments on compensation caps, retention benefits, and rollback policies help explain the roles of the mechanism’s components.

Critical weaknesses
  1. The central feasibility claim is not sufficiently established. Appendix A.2 moves directly from finiteness of the candidate rollback set to nonemptiness of the feasible set without showing that at least one candidate necessarily satisfies all leaving clients’ compensation caps. Algorithm 2 also explicitly includes an infeasibility branch. This gap further affects the feasibility argument in Lemma 3.3, which relies on this claim.

  2. The paper assumes that clients truthfully report their private privacy costs, although these reports directly affect compensation and retention decisions and may therefore invite strategic reporting. For a paper involving incentive design, this assumption limits the mechanism’s practical applicability.

  3. The relationship between the proposed mechanism and the joint optimization problem P1 remains unclear. Stage 1 retains every client whose compensation is feasible without comparing the retention benefit against its payment, so it is unclear whether this rule maximizes server utility.

  4. The formulation imposes individual compensation caps, which is inconsistent with the hard aggregate budget described in the abstract.

  5. FedCARR and SIFU/FedEraser use different forgotten-client sets in the experiments, making it difficult to isolate the advantages of the mechanism or the unlearning procedure from the end-to-end results.

Questions for authors
  1. Could the authors provide explicit conditions guaranteeing that the feasible rollback set is nonempty and explain how these conditions are ensured by the system model?

  2. Under what conditions is the Stage 1 rule of retaining every compensation-feasible client optimal for P1?

  3. Please clearly distinguish Stage 2 optimality, existence of a global optimum, and global optimality of the proposed algorithm.

  4. Does the paper consider individual compensation caps or a shared aggregate payment budget? Please align the abstract, problem formulation, and related claims.

  5. Could FedCARR, SIFU, and FedEraser be compared using the same forgotten-client set and the same recovery-training conditions?

Reviewer_gGjZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

FedCARR proposes a two-stage mechanism for batched federated-unlearning requests: requesters whose modeled retention cost fits a contribution-scaled individual cap are retained and compensated, after which a common rollback checkpoint is selected for the remaining clients. The server objective combines a hand-designed model-value proxy, retained contribution weight, and compensation, and the selected checkpoint is followed by gradient-ascent unlearning and recovery training. Experiments on CIFAR-10, CIFAR-100, and Fashion-MNIST report utility, accuracy, membership-inference, and timing measurements. Although connecting retention decisions with executable rollback checkpoints is sensible, the advertised budget is absent from the formulation, a central feasibility lemma is false, and neither the economic mechanism nor the unlearning evaluation supports the main claims.

Critical strengths
  1. The paper identifies a practically relevant interaction: retaining some requesters changes both the residual deletion set and the set of checkpoints usable for common rollback.
  2. The two-stage procedure is simple and exposes operational quantities—payments, retained contribution, rollback round, and deletion cost—that are often omitted from unlearning evaluations.
Critical weaknesses
  1. Lemma 3.2 is false. Its proof establishes that the candidate set is finite and nonempty and then simply asserts 1≤∣Tfeasible∣<∞1\le|\mathcal T_{\mathrm{feasible}}|<\infty. A constrained subset of a nonempty finite set can be empty. For example, with one leaving client, tir=0t_i^r=0, tiideal=5t_i^{\mathrm{ideal}}=5, Tmax⁡=10T_{\max}=10, ci=1c_i=1, and qiB=1q_iB=1, Stage 1 makes the client leave because Li(10)=10>1L_i(10)=10>1, while the sole rollback candidate has Li(5)=5>1L_i(5)=5>1, so Tfeasible=∅\mathcal T_{\mathrm{feasible}}=\varnothing. Algorithm 2 even includes an infeasible return, contradicting the lemma. The experiments' observation that emptiness did not occur cannot repair the theorem, and subsequent existence claims need an additional assumption or fallback.

  2. The paper does not formulate the advertised hard budget, and FedCARR is not shown to solve P1. The abstract repeatedly invokes a limited global compensation budget, but the model explicitly imposes only per-client caps ri≤qiBr_i\le q_iB, “rather than ... a global payment budget”; there is no constraint such as ∑iri≤B\sum_i r_i\le B. Aggregate expenditure can therefore grow with the number of requesters. Moreover, Stage 1 retains every individually feasible client without comparing its payment with its marginal value to the server, so P1 can prefer unlearning a client whom Stage 1 retains. The paper proves that a finite optimization problem has an optimum and that Stage 2 searches rollback candidates after fixing Stage 1, not that the two-stage mechanism finds or approximates that optimum.

  3. The incentive model assumes away its central strategic problem. The privately known cost cic_i is assumed to be truthfully reported, with no incentive-compatibility or strategy-proofness result. Update-norm contribution qiq_i is neither a causal marginal-value measure nor robust to strategic scaling, and accepted clients are paid exactly their modeled loss, receive zero surplus, and are assumed to accept when indifferent. Thus, the individual-rationality result holds only under truthful declarations and favorable tie-breaking; it is not a credible mechanism-design guarantee.

  4. The unlearning evaluation is not apples-to-apples and provides no deletion guarantee. FedCARR forgets only the residual set F=IleaveF=I_{\mathrm{leave}}, while SIFU and FedEraser are asked to forget the entire request set QQ; their test, retained, forgotten, attack, and runtime results therefore concern different tasks. Forgotten-client training accuracy and one maximum-softmax attack are insufficient evidence that influence has been removed. The reconstruction pipeline begins from checkpoints trained on all clients and applies gradient ascent on forgotten clients' data, but the paper does not explain how the server retains access to those data after consent is revoked. Unlike certified approaches such as SIFU, the operator has no indistinguishability or retraining-equivalence guarantee.

  5. The novelty and empirical utility evidence are weak. Ding et al. already study paying valuable users who request departure to remain through a richer multi-stage incentive game, while Lin et al., WWW 2024 jointly study incentives and dynamic client selection for federated unlearning. Relative to this literature, the common-checkpoint coupling is an incremental addition handled by finite enumeration. The saturating value function G(T)G(T), conversion coefficient, contribution score, and privacy costs are not empirically calibrated or fully specified; positive utility under the paper's own constructed proxy is therefore not strong external validation.

Questions for authors
  1. What assumption guarantees Tfeasible≠∅\mathcal T_{\mathrm{feasible}}\neq\varnothing, and how do the mechanism and theorems change when Algorithm 2 returns INFEASIBLE?
  2. Is BB intended as a global budget or merely a cap coefficient? If it is a budget, where is the aggregate constraint and how is the coupled client-selection problem solved?
  3. Can the authors prove that the Stage-1 threshold rule and Stage-2 search jointly solve or approximate P1?
  4. What makes truthful reporting of cic_i optimal, and how is manipulation of the update-norm contribution score prevented?
  5. Can all methods be evaluated on the identical forget set, with a matched full-retraining reference, stronger influence/privacy tests, and a clear policy for access to withdrawn clients' data?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.